Plant disease classification in the wild using vision transformers and mixture of experts.

Abstract

Plant disease classification using deep learning techniques has shown promising results, especially when models are trained on high-quality images. However, these models often suffer from a significant drop in their accuracies when tested in real-world agricultural settings. In the wild, models encounter images that are significantly different from the training data in aspects like lighting conditions, capturing conditions, image resolution, and the severity of disease. This discrepancy between the training images and images in-the-wild conditions poses a major challenge for deploying these models in agricultural settings. In this paper, we present a novel approach to address this issue by combining a Vision Transformer backbone with a Mixture of Experts, where multiple expert models are trained to specialize in different aspects of the input data, and a gating mechanism is implemented to select the most relevant experts for each input. The use of Mixture of Experts allows the model to dynamically allocate specialized experts to different types of input data, improving model performance across diverse image conditions. The approach significantly improves performance on diverse datasets that contain a range of image capturing conditions and disease severities. Furthermore, the model incorporates entropy regularization and orthogonal regularization, aiming to enhance the robustness and generalization capabilities. Experimental results demonstrate that the proposed model achieved a 20% improvement in accuracy compared to Vision Transformer (ViT). Furthermore, it demonstrated a 68% accuracy on cross-domain datasets like PlantVillage to PlantDoc, surpassing baseline models such as InceptionV3 and EfficientNet. This highlights the potential of our model for effective deployment in dynamic agricultural environments.

Plant phenotyping relevance

野外画像から植物病害の状態・重症度を推定する分類手法を開発し、異なるデータセットや撮影条件で性能検証しているため、植物フェノタイピング手法が中心です。

Code and data availability

The paper's authors publicly release implementation details and trained models on GitHub, and the study analyzes two public plant disease image datasets (PlantVillage and PlantDoc) hosted on Kaggle, all directly used for this paper's phenotyping/classification analysis.